A multi-task Bayesian Hierarchical Gated Transformer for predicting postoperative vault after Implantable Collamer Lens implantation

Abstract Accurate prediction of postoperative vault following Implantable Collamer Lens (ICL) implantation is important for surgical safety, as an inappropriate vault may contribute to complications such as angle closure, cataract formation, and elevated intraocular pressure. Despite advancements in preoperative biometric measurements, existing Machine Learning (ML) approaches often have difficulty accounting for the multifactorial dependencies and inherent uncertainties in clinical data. To address these issues, this study proposed a Multi-task Bayesian Hierarchical Gated Transformer (MBHGT). The model was developed and evaluated on a retrospective dataset of 324 eyes (162 patients) using 16 preoperative parameters. MBHGT combined a Transformer backbone with a Bayesian approximation to quantify predictive uncertainty, a hierarchical gated module for dynamic feature interaction, and a multi-task framework for joint regression and risk stratification. In comparative experiments, MBHGT achieved a Mean Absolute Error (MAE) of 138.01 μm in continuous vault prediction, outperforming the compared baselines. In a 3-class vault classification task (low, normal, high), MBHGT attained a macro-AUC of 0.86 and a macro-F1 score of 0.68, mitigating the majority-class bias commonly observed in conventional models. Quantitative uncertainty evaluation suggested reasonable predictive calibration in both the regression and classification tasks. SHAP-based interpretability analysis quantified individual feature contributions and identified Horizontal Sulcus-to-Sulcus distance (STS), ICL size, and Lens Thickness (LT) as the dominant predictors in our dataset. By combining predictive performance with interpretability, MBHGT shows promise as a decision-support tool for postoperative vault prediction; however, multi-center external and prospective validation is required before clinical application.